The serving section documented only vLLM's --reasoning-parser qwen3. Compact is the llama.cpp release, and llama.cpp users had no documented equivalent for separating reasoning, which is required for multi-turn agent use.
Set mtp_num_hidden_layers to 0 to match the released checkpoint
config.json declared mtp_num_hidden_layers=1, but the released checkpoint contains no MTP tensors: model.safetensors.index.json has 31666 tensors, zero matching mtp/nextn, and decoder layers 0-39 only against num_hidden_layers=40.
Because the config advertised an extra multi-token-prediction module, stock GGUF conversion looked for blk.40 and hard-failed. Users had to pass --no-mtp to convert at all. Setting the count to 0 makes the config describe what actually ships, so conversion works by default.
Weights are unchanged; this is a config-only correction.
